Home/
Part XV — Troubleshooting, Checklists, and Reference/45. Common Failure Modes (And Fix Recipes)
45. Common Failure Modes (And Fix Recipes)
Overview and links for this section of the guide.
On this page
Failure Mode Overview
LLMs fail in predictable ways. Learn the patterns, apply the fixes:
┌─────────────────────────────────────────────────────────────────┐
│ COMMON FAILURE MODES │
├─────────────────────────────────────────────────────────────────┤
│ │
│ IGNORES INSTRUCTIONS │
│ └─ Fix: Structured prompts, examples, repeat key rules │
│ │
│ MESSY OUTPUT │
│ └─ Fix: Schema enforcement, output format examples │
│ │
│ CONFIDENT BUT WRONG │
│ └─ Fix: RAG, self-consistency, uncertainty markers │
│ │
│ INCONSISTENT │
│ └─ Fix: Lower temperature, few-shot examples, testing │
│ │
│ TOO SLOW/EXPENSIVE │
│ └─ Fix: Smaller model, caching, prompt compression │
│ │
└─────────────────────────────────────────────────────────────────┘Quick Fix Guide
| Symptom | First Try | Then Try |
|---|---|---|
| Ignores rules | Move rules to end of prompt | Add examples showing rule-following |
| Wrong format | Use response_mime_type: 'application/json' | Add schema to prompt |
| Makes stuff up | Add "If unsure, say so" | Use RAG with citations |
| Different each time | Set temperature=0 | Add few-shot examples |
| Too slow | Use Flash instead of Pro | Reduce context size |
Where to go next
Explore next
45. Common Failure Modes (And Fix Recipes) sub-sections
5 pages